I’ve been using Neovim since college. I love the extreme customizability, the efficiency of using only a keyboard, and being able to write configurations in Lua. pi is that for coding scaffolds. I started with Claude Code but quickly grew frustrated with the limited vim-mode (though, I am immensely grateful that it was there in the first place and deeply upset when I realized Codex doesn’t have one), the lack of visibility into what exactly went into the context and the frequent, often show-stopping, bugs.

Pi is modularly designed, written in Typescript, and designed to be mucked around with. I quickly was able to spin up a sophisticated vim mode (pi-vim), a subagents extension (pi-subagent) and an OAuth provider (pi-hawk) for Okta, which is what we use at METR for internal proxies. The SDK is really convenient to design against, especially when you want a unified interface with multi-provider AI company backends. The ethos is simplicity, extensibility, and meticulous context management, and eschews things like plan or team mode (just talk to the model).

It’s well-maintained, frequently updated, and the developer is delightfully curmudgeonly (meant with all the respect in the world and in the best way). Go try it out!